Jonathan Shafer

Also known as יונתן שפר

Jonathan Shafer is a theoretical computer scientist who studies how much data machines need to learn and how their results can be verified and secured. After an MSc at Tel Aviv University, he took part in Simons Institute programs as a PhD student at UC Berkeley and published his dissertation, advised by Shafi Goldwasser, in 2024. He coauthored a 2021 paper on interactive proofs for verifying machine learning, a 2023 paper on transductive online learning, and a 2025 STOC paper on removing backdoors from machine learning models. In 2025, his paper Optimal Mistake Bounds for Transductive Online Learning, written with Zachary Chase, Steve Hanneke and Shay Moran, was a NeurIPS Best Paper runner-up. He is a postdoctoral associate at MIT, hosted by Vinod Vaikuntanathan.

Now at: MIT CSAIL

Timeline

2015–2017

Teaching at Tel Aviv University

At Tel Aviv University, he earns an MSc with a thesis on the connections between learning theory and information theory, supervised by Amir Shpilka and Amir Yehudayoff. He serves as a teaching assistant in computational linguistics courses in 2015 and 2016, and in 2017 he teaches Algorithms for Computational Linguistics in place of a professor on sabbatical.
2018–2019

Visiting the Simons Institute

While a PhD student in the theoretical computer science group at UC Berkeley, he joins the Simons Institute programs The Brain and Computation and Lower Bounds in Computational Complexity in 2018, and Foundations of Deep Learning in 2019, as a visiting graduate student.
2021

Interactive proofs for machine learning

In 2021, he coauthors Interactive Proofs for Verifying Machine Learning with Shafi Goldwasser, Guy Rothblum and Amir Yehudayoff at the ITCS conference. The paper studies interactive proof systems in which a verifier, using labeled data and an untrusted prover, checks that a hypothesis is approximately correct while using less data than learning requires.
2023

A trichotomy for transductive learning

In 2023, he coauthors A Trichotomy for Transductive Online Learning with Steve Hanneke and Shay Moran at NeurIPS. The paper proves that the minimal number of learner mistakes grows as n, Θ(log n) or Θ(1), depending on the VC dimension and the Littlestone dimension.
January 2024

A Berkeley dissertation on learning theory

In January 2024, his dissertation Topics in the Theory of Learning appears as UC Berkeley EECS technical report UCB/EECS-2024-2, with Shafi Goldwasser as advisor. Its three parts cover how much data supervised learning needs, whether results from an untrusted machine learning provider can be verified cheaply, and a taxonomy of stability definitions in learning theory.
2025

Backdoor removal without detection

In 2025, he coauthors Oblivious Defense in ML Models: Backdoor Removal without Detection with Shafi Goldwasser, Neekon Vafa and Vinod Vaikuntanathan, which appears at STOC 2025. The paper shows that backdoors in machine learning models, even undetectable ones, can sometimes be mitigated or removed with black-box techniques inspired by random self-reducibility.
2025

NeurIPS Best Paper runner-up

In 2025, he coauthors Optimal Mistake Bounds for Transductive Online Learning with Zachary Chase, Steve Hanneke and Shay Moran, which receives a NeurIPS 2025 Best Paper runner-up award. The paper resolves a 30-year-old open problem on the power of unlabeled data and establishes a quadratic gap between transductive and standard online learning.
As of October 2026

Postdoctoral associate at MIT

He is a postdoctoral associate at MIT's Computer Science and Artificial Intelligence Laboratory, hosted by Vinod Vaikuntanathan, and co-organizes the MIT ML+Crypto Seminar. He works in theoretical computer science, with a focus on computational and statistical learning theory.